It was a Tuesday morning standup when one of my senior analysts said something that stopped the meeting cold.
“I’ve been using ChatGPT every day for eight months,” she said. “But I still don't actually trust it.”
She wasn’t complaining. She wasn’t asking for help. She was just stating a fact, the way you’d say “I drive to work every day but I still don’t fully trust the highway.” The room went quiet for a second. Then three other people nodded.
That moment stuck with me. Because this week, two data points landed in the same 72-hour window that made that conversation feel like the most important thing I’ve heard all year.
First: ChatGPT crossed 1 billion monthly active users in June 2026. The fastest app in history to reach that milestone. Faster than TikTok, faster than Instagram, faster than anything. One billion people.
Second: Stanford HAI dropped its 2026 AI Index — 423 pages, nine years of independent research — and buried in the public sentiment chapter was this: only 23% of the general public believes AI will have a positive impact on jobs. Among AI experts? 73%.
Let that gap sink in.
One billion users. And only one in four of them actually believes in what they’re using.
I’ve been leading data science, analytics, and platform engineering teams for over a decade. I’ve shipped AI systems into production, watched them fail, rebuilt them, and shipped them again. And I think I finally understand what’s creating this gap — and why most of the industry refuses to talk about it honestly.
The Number That Should Embarrass Every AI Leader
The Stanford report documents something I’ve been watching from the inside for years, but seeing it quantified is different.
73% of AI experts are optimistic about AI’s impact on the job market. 23% of the general public agrees. That’s not a communications gap. That’s not a media literacy problem. That’s a 50-point chasm between the people building these systems and the people living with them.
And it’s getting worse, not better.
Among Gen Z — the generation that grew up with smartphones, that should theoretically be the most comfortable with AI — the share who describe themselves as excited about AI dropped from 36% in 2025 to 22% in 2026. The proportion feeling angry rose from 22% to 31%.
These are people who use AI daily. They’re not afraid of it the way their grandparents might be. They’re angry at it. That’s a fundamentally different signal, and I don’t think we’re reading it correctly.
What 10 Years in Production Rooms Taught Me About This Gap
Here’s the uncomfortable truth I’ve learned running data teams through multiple AI waves: the trust deficit isn’t irrational. It’s the correct response to how we’ve been deploying these systems.
Think about what AI actually looks like from the outside, from the perspective of someone who isn’t in the room where it gets built.
You get a recommendation algorithm that shows you products you already bought. You submit a resume and never hear back because an ATS scored you out before a human saw your name. Your loan application gets denied with a reason you can’t appeal. Your streaming service shows you content based on “what you like” that somehow never matches what you actually like.
These aren’t edge cases. These are the dominant experiences most people have with AI in the real world. And they’re produced by systems built by people who optimized for metrics — engagement rate, processing speed, model accuracy — without ever asking whether the output was trustworthy, explainable, or fair from the perspective of the person receiving it.
I’ve been in those rooms. I’ve made those tradeoffs. And looking at the Stanford numbers now, I think we underestimated the cumulative cost.
The Part Where My Own Team Gets Uncomfortable
About eighteen months ago, we were building a churn prediction model for a business unit I support. The model was good. 87% accuracy on holdout. The business unit head was thrilled.
Then one of my analysts — a junior one, six months in — asked a question that we hadn’t thought to ask: “What does the person getting flagged as ‘likely to churn’ actually experience after that prediction?”
We pulled the thread. What we found was that the downstream intervention — an automated re-engagement sequence triggered by the model — was perceived by customers as intrusive and slightly creepy. Several had written in to complain that we seemed to “know” things about them. The model was right about churn risk. But the intervention it triggered was eroding trust faster than it was preserving it.
We rebuilt the experience from the customer’s perspective. Accuracy dropped two points. Customer trust scores on the intervention went up 31%.
Two accuracy points. Thirty-one trust points.
I think about that ratio a lot.
The Transparency Problem Nobody Wants to Claim
The Stanford Index caught something else that most coverage has missed: more than 80 of the 95 most notable AI models released in 2025 were released without their training code. Google, Anthropic, and OpenAI have all stopped disclosing their latest models’ dataset sizes and training duration.
The models are getting more capable. They’re also getting more opaque.
From a competitive standpoint, I understand this completely. From a public trust standpoint, it’s catastrophic. We’re asking people to trust systems we’ve deliberately made impossible to audit. And then we’re surprised when only 23% of them do.
There’s a version of this conversation where I argue that interpretability will solve it, that XAI tools will eventually make these systems legible enough to earn public trust. I used to believe that. I’m less sure now.
I think the actual fix is harder: it’s redesigning systems with trust as a primary constraint, not an afterthought. Before asking “will this model be accurate,” asking “will the person on the receiving end of this output understand what happened and have a path to contest it?” It adds friction. It sometimes reduces headline metrics. It is almost never what you optimize for when you’re moving fast and trying to ship.
What I’m Actually Changing in My Practice
After digesting the Stanford numbers this week, I did something I don’t usually do: I put it on the agenda for my team’s next sprint planning.
Not as a philosophical discussion. As a design constraint.
We’re introducing what I’m calling a trust audit as part of our model deployment checklist. For every model going into production, we’re requiring the team to answer three questions before shipping:
Can the person most affected by this output understand why they received it? Can they appeal it? Have we tested whether the downstream experience of the output matches the intent of the model?
It’s not enough to ship a model that works. We have to ship a model that’s trustworthy — and those are not the same thing, and the industry has spent five years pretending they are.
The Stanford report also noted that US private AI investment hit $285.9 billion in 2025, more than 23 times China’s $12.4 billion. We are spending at historic scale. And we are generating historic distrust simultaneously. That is a structural problem, not a messaging problem.
The Thing My Analyst Was Really Telling Me
When she said “I use it every day but I still don’t actually trust it,” she wasn’t describing a failure of understanding. She was describing a rational response to a system that’s been designed to be used, not to be trusted.
She’s one of the most technically sophisticated people on my team. If she feels that way, I have a hard time blaming the billion people who feel it about products built with far less care.
One billion users is an extraordinary number. I don’t want to minimize what it represents about AI’s genuine utility. People are using these tools because they are useful.
But trust isn’t built by being used. Trust is built by being reliable, transparent, and fair across millions of interactions with people who can’t see inside the system producing those interactions. That’s the work we haven’t done yet. And the Stanford data is the receipt.
The question I keep coming back to: at what point does the trust gap become the capability gap? When does “people don’t trust AI” start showing up not just in sentiment surveys but in adoption ceilings, regulatory crackdowns, and genuine public backlash that slows the whole project down?
I think we’re closer to that moment than most of the people building these systems want to admit.
What’s your read on this? Have you had moments in production where you realized you’d optimized for the wrong thing? I’m genuinely curious whether other data leaders are factoring trust as a design constraint, or whether it’s still being treated as a communications problem to be solved downstream.
🌐 Stay Connected 🌐
If you enjoyed this post, tap the 👏 and hit “Follow” to stay updated with my latest stories on data, tech, and reflections.
👥 Follow me on LinkedIn 🧠 Let’s exchange ideas 💡 Share your thoughts & perspectives
Comments
Loading comments…